is there something wrong with hard negative mining?

Question

is there something wrong with hard negative mining?

Opened this issue 6 years ago · 3 comments

the implement of hard negative mining actually decrease the anchor number so when positive number = 0,negative number will be = 0 too

Answer 1 · 2018-11-19T01:24:14.000Z

positive number != 0, because https://github.com/XiaXuehai/faceboxes/blob/master/dataset.py#L131.
There is an imbalance between the positive and negative examples if you do not use hard negative mining.

Answer 2 · 2018-11-19T07:39:52.000Z

positive number != 0, because https://github.com/XiaXuehai/faceboxes/blob/master/dataset.py#L131.
There is an imbalance between the positive and negative examples if you do not use hard negative mining.

yes,the box in the random crop will no be 0 but the positive anchor will be 0 because of the IOU

Answer 3 · 2018-11-19T10:00:32.000Z

When encoding, we set the max IOU to be the positive anchor.